Causal Analysis for Understanding Vehicle Behavior Affected by Multiple Factors

Nowadays, autonomous vehicles rely on machine learning approaches to achieve a higher level of automation and improve driving performance. Machine learning approaches, in turn, learn vehicle behavior through the observation of other drivers and the environment, the latter being a key factor that influences vehicle behavior. However, the amount of influence of the environment on vehicle behavior cannot be grasped by uniformly analyzing the entire driving route. Thus, this paper proposes a range-based causal analysis to inspect the effects that specific locations have on vehicle behavior. In other words, to use range or distance from a target to concentrate the analysis of vehicle behavior on a particular location. To interpret vehicle behavior, we employed a semantic primitive motion clustering that combines Hierarchical clustering and HDP-HSMM. Common causes for vehicle behavior change were then extracted by tracking the traffic participants, and the environment was segmented by region and ranges of each region, comparing three different types of regions: crosswalks, traffic lights, and entrances. The results show valuable insight into vehicle behavior by highlighting key primitive motions that take place only in a particular region depending on the range of proximity. Moreover, a comparison of the three studied regions shows that they present singular patterns of behavior if analyzed by range. For instance, vehicles heading to an entrance tend to drastically reduce velocity when being close to an entrance and pedestrians. The same does not happen in proximity to other vehicles, in which case the vehicle comes to a complete stop. Moreover, it was observed that the vehicle behavior changes according to each of the three traffic regions considered in the study. This insight into the key primitives obtained using the proposed approach can be key to understanding different scenarios of interaction between the environment and autonomous vehicles.